A couple of short prompts using Lean for formal verification resulted in 16 PRs fixing bugs and race conditions in the Claude Agent SDK. bcherny also combines Lean with TLA+ for data flow and concurrency analysis, and notes you don't need to know either language yourself since the model handles both fluently.
Opus 5.5 launched with a significant price reduction driven by efficiency gains, plus increased rate limits, making the kind of multi-agent formal verification workflows being demoed accessible to individual developers.
Switching a PR review bot from Opus 5 to Opus 5.5 cut costs by 50%, reduced tool calls by 34%, and slashed execution time by 62%, with no quality loss. An Anthropic team member flagged this as impressive.
Responding to a Formal Verification researcher who found earlier models bad at generating high-level system properties, the team member said Opus 5.5 feels much better and invited the researcher to test it themselves. The context came from using Opus 5.5 with Lean and TLA+ to formally verify the Claude Agent SDK.
A demo where Opus 5.5 generated a visual that X-rays a fictional house, showing pipes, wires, and ducts, then identified four design mistakes before construction. The framing suggests this could be a useful starting point for an expert builder to build on.
A developer asked Opus 5.5 to explain camera focus by building an interactive lens lab. It ran for 1 hour 26 minutes, cost $25.66, and produced a working simulation where moving the focus ring shifts glass elements and the sharp plane through the scene. bcherny suggested this points toward a new era of personalized educational tools.
A user prompted Opus 5.5 to animate its own existence from day zero, generating every frame as JavaScript brush strokes with no video or image model involved. The Anthropic team member who saw it floated the idea of a full graphic novel next.
The Claude Code team is considering removing plan mode entirely and replacing it with a shift+tab shortcut to adjust effort levels, on the grounds that current models no longer need a separate planning phase. Feedback from plan mode fans is requested.
Following user pushback on removing plan mode, trq212 floats a middle path: keeping plan mode for a subset of models where it may still be necessary, while removing it for more capable ones.
In response to a user who relies on plan mode to preview and refine work before committing, a team member suggested using artifacts for planning as an alternative, or adjusting plan mode to lean on artifacts more.
In response to a user asking whether Opus 5.5 could dynamically shift reasoning effort mid-task based on what comes next, a team member said they are thinking about it. The idea is to get the benefits of effort scaling without requiring users to manage reasoning levels manually.
The ClaudeDevs account published a post detailing how Claude was used to measure, debug, and improve performance on claude.ai, resulting in a 3x speed increase over two weeks. Specific prompts and techniques are included.
The recent speed improvements to claude.ai and the Desktop app came from techniques now detailed in a blog post, aimed at engineers working on performance in their own apps.
A new content format from the team: detailed posts sharing the actual prompts and methods used internally, replicable by other engineers. The inaugural entry covers how claude.ai was made 3x faster in two weeks.
A code-generated Golden Gate Bridge scene built entirely by Opus 5.5 looks comparable to Google Astra's output on the same scene, per testing. The full comparison is in a linked video.
A developer ran four Opus 5.5 agents for 90 minutes to produce what they called a masterpiece, built entirely from scratch with no AI voice API. bcherny called it gorgeous.
A notable property of Opus 5.5 is that it scales gracefully across effort levels: low for speed, xhigh for maximum reasoning, with medium as bcherny's everyday default.
On Claude Code v2.1.280 and later, changing the effort level mid-session no longer busts the prompt cache when using Opus 5.5.
The cache window is 1 hour for usage within a subscription plan and 5 minutes for API key or extra usage. Users can override the timeout by setting "promptCacheTtl": "1h" in settings.json or via the CLAUDE_CODE_PROMPT_CACHE_TTL environment variable, and can also pin the cache.
